Context: despite v0.37.0 shipping async ingest correctly, a user
session spent 3h 48m on two 1-minute PDF ingests because the agent set
up a bash 'until grep ... done' polling loop and kept calling
jobs_status instead of returning control. The ingest daemon was idle
within 2 minutes; the agent burned the rest.
Alexandria's async ingest response was encouraging this with a
'Poll with: jobs_status(...)' line that read like 'poll immediately'.
* mcp/tools/write_tool.py: replace the 'Poll with' suggestion with an
explicit 'return control to the user, do NOT spin-poll, do NOT set
up a bash polling loop' instruction in the async-ingest response.
* alexandria/skill/SKILL.md: add an Anti-patterns section listing
exactly the behaviours observed (bash polling loops, serial
jobs_status calls in one turn, 'proving it works' via bash before
retrying, waiting on wait_s=60 for batches). Strengthens the
batch-ingest workflow to say 'Return control. Stop.' after
enqueueing.
* Bump version to 0.37.2